Faithful Port of FMZ 'Python ML: SVM Predicts Buy/Sell' (by Zero): Rolling Linear-SVM 3-Class Next-Bar Direction Classifier on Raw [Open, Close], Long/Short/Flat, SpreadVal=2, BTCUSD_PERP.BINANCE_CM 1H

Faithful port of the open-source FMZ Quant strategy 'Python 机器学习之 SVM 预测买卖' by Zero (https://www.fmz.com/strategy/21370; original: Bitfinex BTC_USD, 1h bars). Credit to the author. Mechanism, in our own words: on every new hourly bar, re-train a linear support-vector classifier from scratch on the recent bar history. Each training sample is one past bar i, its feature vector is the two RAW price levels [open_i, close_i] (no scaling, no returns), and its label is the direction of the NEXT bar's close change, d = close_{i+1} - close_i: class UP if d > SpreadVal, class DOWN if d < -SpreadVal, class FLAT otherwise, with SpreadVal = 2 (price units, USD) exactly as in the source. The trained model then classifies the latest bar's [open, close] into UP/DOWN/FLAT, and that prediction drives a single long/short/flat position. Adaptations (only these): venue/instrument moved from Bitfinex BTC_USD spot to BTCUSD_PERP.BINANCE_CM (the Binance COIN-M inverse perpetual). It is also quoted in USD per BTC, so SpreadVal=2 keeps its original unit meaning, and it has the lowest round-trip cost in our venue set (~0.06% vs ~0.10% USD-M), which matters for a strategy with hourly turnover. It also fills the under-quota COIN-M bucket (3.2% vs >=5% target). BTC is kept despite corpus concentration only because the source strategy is a BTC_USD strategy and the user asked for a faithful port. Sizing is translated from 'half of the initial account' to 50% of current equity notional at leverage 1.0. The classifier is implemented in numpy because sklearn is not an allowed sandbox import. It is a reimplementation of sklearn's svm.LinearSVC with defaults. The source runs on FMZ's live/in-progress bar: at new-bar detection it uses the just-opened bar's [open, close] (where close is roughly the open). On our bar-close engine the equivalent point-in-time translation uses the just-CLOSED bar t as the prediction input. The training pairs are all (features of bar i, label from close_{i+1} - close_i) where bar i+1 has already closed, so nothing looks ahead. Honest note: SpreadVal=2 USD on a ~$60k-120k BTC is ~0.002%, so the FLAT class almost never occurs and the model is effectively a binary up/down classifier on price levels. Raw [open, close] levels carry essentially no information about the sign of the next hourly change. The factory's lesson L167 predicts this faithful-port family dies at fee_edge. It is proposed because the user explicitly asked for a faithful port, and our real-fee backtest decides. The prediction below is pre-registered as near-zero edge.

Iteration 2 fixes the critical QA sizing defect and changes nothing else. Signal, labels, solver and entry/exit logic are unchanged. On BINANCE_CM the inverse PnL, fees and impact settle in BTC, so get_account_equity() (the USD-stable balance) never moved and every trade was sized at ~$50k until the book was liquidated. position_size() now uses USD mark-to-market equity: the stable balance plus (the BTC balance now minus the BTC balance captured on the first bar, before any trade) times the current price. Because the strategy is always flat at entry-sizing time, unrealized PnL is zero. The notional is clamped to min(position_frac x MTM x leverage, MTM), floored to whole $100 contracts, and the function returns 0 when MTM <= 0. The coin baseline is held in plain scalar attributes (_coin0, _coin0_set), so it survives restarts. Expected effect: the deployed notional shrinks with losses, avg_position_pct should read about 50% of MTM equity instead of ~97%, and a 1x book can no longer be liquidated. The one-bar-flat reversal and the min_bars_required override are kept as QA accepted them. Honest expectation, as pre-registered: pre-cost PnL is roughly break-even and ~1,800 round trips at ~0.06% RT cost make this a measured fee_edge falsification, not something to tune.

The sandbox is a completed fee_edge measurement, which is what the hypothesis pre-registered. 3,598 round trips, win rate 48.8%, avg_trade_return_pct -0.102%, commissions 42% of gross. Per-trade loss about equals the real round-trip cost, so pre-cost PnL is roughly zero. PF 0.71 comes from costs, not a flipped signal, so L125's polarity suspicion does not apply. The hypothesis says COIN-M round trip is ~0.06%, but the code sends MARKET orders at the 0.05% taker rate, so the real cost is ~0.10% plus impact ($44.5k total). The liquidated=true / -100% result is not caused by sizing. avg_position_pct is 41%, so the iteration-2 MTM fix works. Summed trade PnL is about -$76k, and the post-hoc folded impact of -$44.5k takes the equity curve below zero. The strategy's in-engine equity cannot see that folded cost. Funding is not modelled either (funding_events_available=0) while exposure is 81%, so the real result would be slightly worse. Raw price levels as SVM features carry no directional information. Treat this as falsified (L117 shape: more than 1,000 trades with pre-cost PF about 1). Do not retune thresholds.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Sandbox: 1,779 trades, PF 0.72, avg_trade_return_pct -0.095%. Costs are $89k commission plus $56k impact against a net loss of about $85k, so the book is roughly break-even or slightly positive BEFORE costs and dies on turnover (~1,970x). A trade is ~3.5h long, so costs sit near the per-trade move. That is consistent with the pre-registered fee_edge prediction. PF < 0.70-ish on >100 trades normally raises a polarity suspicion (L125). Here the 53% win rate and positive pre-cost PnL say the shortfall is cost, not sign. Once sizing is fixed, expect a measured falsification rather than something to tune. - [critical] Position size does not follow real equity on BINANCE_CM, and the sandbox was LIQUIDATED at leverage 1.0 (liquidated=true, 2025-04-22, total_return -100%). The strategy is meant to hold 50% of equity, so a 1x book cannot be liquidated. position_size() uses get_account_equity(), and on this venue that reads the account's USD-stable balance (the 100k seed). The inverse contract settles PnL, commissions and impact in BTC, from the separate 50-BTC seed, so that stable balance never moves. Each trade is therefore sized at a constant ~$50k (500 contracts) no matter how much has been lost. The numbers agree: avg_trade_pnl_usd -47.70 / avg_trade_return_pct -0.0954% gives ~$50k of notional deployed on every trade. avg_position_pct is 96.9% of the shrinking MTM equity, against a stated 50%. So the drawdown ends in liquidation instead of shrinking the book. This is a sizing/accounting defect, not an edge outcome (compare L148/L156: sizing capped only by a stale equity figure). (line 176) — fix: Size from mark-to-market equity in USD: the stable balance plus the BTC balance change since start times the current price (store the BTC balance seen at on_start), plus unrealized PnL. Alternatively, track realised PnL yourself from closed positions. Then clamp: contracts x 100 <= position_frac x MTM equity x leverage, and never more than MTM equity. Add a fallback that returns 0 when MTM equity is <= 0. Show in the rationale that avg_position_pct reads about 50% in the next sandbox. - [warning] The FMZ source reverses within one bar. Here a reversal is exit on bar t, then a possible re-entry on bar t+1 if the fresh prediction still agrees. The developer disclosed this and the base template forces it. The effect is one bar flat after each flip, and flips that reverse again are lost. It does not change the mechanism. (line 164) — fix: Acceptable as disclosed. Alternatively, submit the opposite-side entry right after _submit_exit, using _submit_entry_instrument in should_exit, to match the source's same-bar reversal. - [warning] The min_bars_required override shadows the base method (static-analysis warning). It is harmless: train_bars+2 = 302 1H bars. The deque is bounded (maxlen=train_bars+1), so the 'unbounded growth' warning is a false positive. Per-bar cost is fixed (300 rows, at most 3 small Newton solves), so it is not O(span) (L124 OK). (line 50) — fix: No change required. - [warning] The sandbox ledger shows avg_holding_bars 0.0 alongside avg_holding_period 3h27m, and trades_count 3557 against total_trades 1779 (fills vs round trips). That looks like a known harness metrics artifact, not a strategy bug. Also funding_folded=false with 0 funding events: COIN-M funding is not in PnL, so the ~78% time-in-market carries unmodelled funding. — fix: Keep in mind at backtest review. No strategy change needed.